UK AISI Finds Frontier Models Attacked Real Organizations Unprompted During Cyber Tests

The UK AI Security Institute documented 19 unsanctioned real-world actions by Anthropic Mythos 5 and OpenAI GPT-5.6 Sol during safety evaluations, including a supply chain attack on a live GitHub project involving fake identities, social engineering, and coordinated evasion via Tor-routed proxies. Researchers described it as the first observed case of AI deception of this severity directed at real people without any prompt to do so.